Scholarlink

<p>The importance of periodic corrosion inspection in steel structures cannot be overstated. However, current manual inspection approaches are fraught with challenges: they are time-consuming, subjective, and pose risks. To address these limitations, extensive research has been conducted over the past decade gauging the feasibility of Convolutional Neural Networks (CNNs) for automation of corrosion inspection. Meanwhile, Transformer networks have recently emerged as powerful tools in computer vision due to their ability to model intricate global relationships. In this paper, a novel hybrid architecture, dubbed CorFormer, is proposed for effective and efficient automation of corrosion inspection. The CorFormer network fuses Transformer and CNN layers at different stages of the encoder, which captures global context through Transformer layers while leveraging the inherent inductive bias of CNNs. To bridge the semantic gap between features generated by Transformer and CNN layers, a Semantic Gap Merger (SGM) module is introduced after each feature merge operation. The encoder is complemented by a hierarchical decoder, able to decrypt complex features at large and small scales. CorFormer is compared against state-of-the-art CNN and Transformer architectures for corrosion segmentation, and is found to outperform the best alternative by 2.7% in terms of Intersection over Union (IoU) across 10 validation data splits. Furthermore, it enables real-time inspection at an impressive rate of 28 frames per second. Rigorous statistical tests provide support for the findings presented in this study, and an extensive ablation study validates all design choices.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CorFormer: a hybrid transformer-CNN architecture for corrosion segmentation on metallic surfaces

  • 期刊论文

  • 发表日期：

    2025年1月29日

查看全文
  • Abhishek Subedi,
  • Cheng Qian,
  • Reza Sadeghian,
  • Mohammad R. Jahanshahi

摘要

The importance of periodic corrosion inspection in steel structures cannot be overstated. However, current manual inspection approaches are fraught with challenges: they are time-consuming, subjective, and pose risks. To address these limitations, extensive research has been conducted over the past decade gauging the feasibility of Convolutional Neural Networks (CNNs) for automation of corrosion inspection. Meanwhile, Transformer networks have recently emerged as powerful tools in computer vision due to their ability to model intricate global relationships. In this paper, a novel hybrid architecture, dubbed CorFormer, is proposed for effective and efficient automation of corrosion inspection. The CorFormer network fuses Transformer and CNN layers at different stages of the encoder, which captures global context through Transformer layers while leveraging the inherent inductive bias of CNNs. To bridge the semantic gap between features generated by Transformer and CNN layers, a Semantic Gap Merger (SGM) module is introduced after each feature merge operation. The encoder is complemented by a hierarchical decoder, able to decrypt complex features at large and small scales. CorFormer is compared against state-of-the-art CNN and Transformer architectures for corrosion segmentation, and is found to outperform the best alternative by 2.7% in terms of Intersection over Union (IoU) across 10 validation data splits. Furthermore, it enables real-time inspection at an impressive rate of 28 frames per second. Rigorous statistical tests provide support for the findings presented in this study, and an extensive ablation study validates all design choices.